Closed ycendes closed 8 years ago
You can set the detection threshold to >= 18. Lower than this and these two are confused resulting in nonsense values for that gauss fit.
Not sure why the fit gives such nonsense values, negative pixel location, on a bad fit though.
What does your figure show? Is this the source that it is failing on? Also, a couple of source-finder questions - is this for a completely free fit or one of the force-fit to a Gaussian? Also, have you tried changing the deblending settings and the RMS grid size?
Ideally, if the image is plagued by RFI then I think we should reject the image automatically. However, it is concerning that TraP crashes so badly on this image - it should simply be skipping any sources that it fails to fit or gives a NAN result.
Yes these are the sources which cause it to fail. The source extractor is seeing this as one source and trying to fit a gaussian to it. I've attached the mask image of all the sources. This one is the light blue near the top-right. note the red double, looking one near the bottom left is actually two.
The result with detection threshold 18 can be seen here http://banana.transientskp.org/master/vlo_marknewdb_2/image/5497/
Yes it is a free fit, that is no parameters are fixed when fitting. No I haven't changed any of the de-blending settings.
That might be worth trying, however I might be tempted to say that perhaps we should find a way to reject these images, if this is simply caused by bad quality data, instead of putting a lot of effort into solving the source finder problem.
I think I agree with @AntoniaR here, we should first try just skipping these images since running source extractor here would ruin your lightcurve anway. @ycendes knows how to do this, but didn't get to it yet.
Still the sourcefinder shouldn't crash. @mkuiack tried to fix this by just continuing on the described exception, but then the sourcefinder logic starts to return invalid data which will crash TraP later on.
Yvette asked me to rerun /scratch/fhuizing/aartfaac/results/shower/s256/ images with the latest master and the same job_params to see if I would reproduce her crash. So I made a new project and a new job, then copied the job_params.cfg and images_to_process.py files, then cloned /tkp from github did and ran TraP, the results are in http://banana.transientskp.org/master/vlo_mark_db_2/dataset/1/ It crashed on image 174 with the same error as above.
I then when back to my other directory and ran it, with the same job_params.cfg and images_to_process.py, but with the tkp files that I'd previously been messing around with and it ran all the way though again. The results are in http://banana.transientskp.org/master/vlo_mark_db_1/dataset/11/
In the run that worked the dataset is added to a database which contains Peeyush's images. And I notice that the image id's reflect that, so they're ~6100 - ~6500 rather than 0-369. Not sure what else would be affected by mixing datasets in a database? Maybe sky coordinates? It still doesn't make sense if the problem is a bad Gaussian fit to a blended source in a high noise image though, so I don't know...
To answer my own question, yes this does result in a shift in the measured coordinates for a source. For example take the same image from both runs: http://banana.transientskp.org/master/vlo_mark_db_2/image/161/ and http://banana.transientskp.org/master/vlo_mark_db_1/image/6306/ then look at its Running Catalogue entry and you can see difference in the position and Light curve. Whether these are within know error bars though I don't know. Dec +- 4'? that seems high!
@mkuiack what runningcatalog are you referring to? When I look at the two images everything looks the same.
Ok I get your point. I think the difference in positions can be explained by the unfinished run. The unfinished run got to about 50%. Since we are blind fitting on a lot of noise I think there is a lot of 'drifting' happening.
so this is actually a duplicate of this issue reported by @hsuyeep right?
So yes, this is a duplicate. I'm closing this issue, please reopen if you think otherwise.
Running through Folkert's latest batch of images, ie at /scratch/fhuizing/aartfaac/results/shower/s256/, I was getting the following:
RuntimeError: Spatial position is not a number
When running in serial mode, here is the full error including the image where it fails (which, upon inspection, is an image that is plagued with RFI)-
I am currently working in ycendes@struis:/scratch/ycendes/aartfaac/cendesrfitest